Separation of speech sources in under-determined case using SCA and time-frequency methods

This paper presents a new algorithm for blind source separation (BSS) of instantaneous speech mixtures in under-determined case. A demixing algorithm which exploits the sparsity of speech signals in the short time Fourier transform (STFT) domain is proposed. This algorithm combines the modified k-means clustering procedure involved in the line orientation separation technique (LOST) with Smoothed l⁰-norm minimization (SL0) method. First procedure along with a transformation into a sparse domain tries to estimate the mixing matrix, and the second method tries to extract the sources from the mixtures. Simulation results are presented and compared to the degenerate unmixing estimation technique (DUET) and LOST algorithms. It is shown in this article that improvements are achieved in two cases. One is the quality of source extraction when the number of mixtures is increased, and second is the speed of source separation when compared to the LOST algorithm. This speed enhancement is about 3 to 10 times comparing to the LOST algorithm. We called the proposed algorithm SL0-LOST.